Index: head/usr.sbin/bhyve/pci_virtio_block.c =================================================================== --- head/usr.sbin/bhyve/pci_virtio_block.c (revision 280724) +++ head/usr.sbin/bhyve/pci_virtio_block.c (revision 280725) @@ -1,410 +1,408 @@ /*- * Copyright (c) 2011 NetApp, Inc. * All rights reserved. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY NETAPP, INC ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL NETAPP, INC OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * $FreeBSD$ */ #include __FBSDID("$FreeBSD$"); #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include "bhyverun.h" #include "pci_emul.h" #include "virtio.h" #include "block_if.h" #define VTBLK_RINGSZ 64 #define VTBLK_S_OK 0 #define VTBLK_S_IOERR 1 #define VTBLK_S_UNSUPP 2 #define VTBLK_BLK_ID_BYTES 20 /* Capability bits */ #define VTBLK_F_SEG_MAX (1 << 2) /* Maximum request segments */ #define VTBLK_F_BLK_SIZE (1 << 6) /* cfg block size valid */ #define VTBLK_F_FLUSH (1 << 9) /* Cache flush support */ #define VTBLK_F_TOPOLOGY (1 << 10) /* Optimal I/O alignment */ /* * Host capabilities */ #define VTBLK_S_HOSTCAPS \ ( VTBLK_F_SEG_MAX | \ VTBLK_F_BLK_SIZE | \ VTBLK_F_FLUSH | \ VTBLK_F_TOPOLOGY | \ VIRTIO_RING_F_INDIRECT_DESC ) /* indirect descriptors */ /* * Config space "registers" */ struct vtblk_config { uint64_t vbc_capacity; uint32_t vbc_size_max; uint32_t vbc_seg_max; struct { uint16_t cylinders; uint8_t heads; uint8_t sectors; } vbc_geometry; uint32_t vbc_blk_size; struct { uint8_t physical_block_exp; uint8_t alignment_offset; uint16_t min_io_size; uint32_t opt_io_size; } vbc_topology; uint8_t vbc_writeback; } __packed; /* * Fixed-size block header */ struct virtio_blk_hdr { #define VBH_OP_READ 0 #define VBH_OP_WRITE 1 #define VBH_OP_FLUSH 4 #define VBH_OP_FLUSH_OUT 5 #define VBH_OP_IDENT 8 #define VBH_FLAG_BARRIER 0x80000000 /* OR'ed into vbh_type */ uint32_t vbh_type; uint32_t vbh_ioprio; uint64_t vbh_sector; } __packed; /* * Debug printf */ static int pci_vtblk_debug; #define DPRINTF(params) if (pci_vtblk_debug) printf params #define WPRINTF(params) printf params struct pci_vtblk_ioreq { struct blockif_req io_req; struct pci_vtblk_softc *io_sc; uint8_t *io_status; uint16_t io_idx; }; /* * Per-device softc */ struct pci_vtblk_softc { struct virtio_softc vbsc_vs; pthread_mutex_t vsc_mtx; struct vqueue_info vbsc_vq; struct vtblk_config vbsc_cfg; struct blockif_ctxt *bc; char vbsc_ident[VTBLK_BLK_ID_BYTES]; struct pci_vtblk_ioreq vbsc_ios[VTBLK_RINGSZ]; }; static void pci_vtblk_reset(void *); static void pci_vtblk_notify(void *, struct vqueue_info *); static int pci_vtblk_cfgread(void *, int, int, uint32_t *); static int pci_vtblk_cfgwrite(void *, int, int, uint32_t); static struct virtio_consts vtblk_vi_consts = { "vtblk", /* our name */ 1, /* we support 1 virtqueue */ sizeof(struct vtblk_config), /* config reg size */ pci_vtblk_reset, /* reset */ pci_vtblk_notify, /* device-wide qnotify */ pci_vtblk_cfgread, /* read PCI config */ pci_vtblk_cfgwrite, /* write PCI config */ NULL, /* apply negotiated features */ VTBLK_S_HOSTCAPS, /* our capabilities */ }; static void pci_vtblk_reset(void *vsc) { struct pci_vtblk_softc *sc = vsc; DPRINTF(("vtblk: device reset requested !\n")); vi_reset_dev(&sc->vbsc_vs); } static void pci_vtblk_done(struct blockif_req *br, int err) { struct pci_vtblk_ioreq *io = br->br_param; struct pci_vtblk_softc *sc = io->io_sc; /* convert errno into a virtio block error return */ if (err == EOPNOTSUPP || err == ENOSYS) *io->io_status = VTBLK_S_UNSUPP; else if (err != 0) *io->io_status = VTBLK_S_IOERR; else *io->io_status = VTBLK_S_OK; /* * Return the descriptor back to the host. * We wrote 1 byte (our status) to host. */ pthread_mutex_lock(&sc->vsc_mtx); vq_relchain(&sc->vbsc_vq, io->io_idx, 1); vq_endchains(&sc->vbsc_vq, 0); pthread_mutex_unlock(&sc->vsc_mtx); } static void pci_vtblk_proc(struct pci_vtblk_softc *sc, struct vqueue_info *vq) { struct virtio_blk_hdr *vbh; struct pci_vtblk_ioreq *io; int i, n; int err; int iolen; int writeop, type; off_t offset; struct iovec iov[BLOCKIF_IOV_MAX + 2]; uint16_t idx, flags[BLOCKIF_IOV_MAX + 2]; n = vq_getchain(vq, &idx, iov, BLOCKIF_IOV_MAX + 2, flags); /* * The first descriptor will be the read-only fixed header, * and the last is for status (hence +2 above and below). * The remaining iov's are the actual data I/O vectors. * * XXX - note - this fails on crash dump, which does a * VIRTIO_BLK_T_FLUSH with a zero transfer length */ assert(n >= 2 && n <= BLOCKIF_IOV_MAX + 2); io = &sc->vbsc_ios[idx]; assert((flags[0] & VRING_DESC_F_WRITE) == 0); assert(iov[0].iov_len == sizeof(struct virtio_blk_hdr)); vbh = iov[0].iov_base; memcpy(&io->io_req.br_iov, &iov[1], sizeof(struct iovec) * (n - 2)); io->io_req.br_iovcnt = n - 2; io->io_req.br_offset = vbh->vbh_sector * DEV_BSIZE; io->io_status = iov[--n].iov_base; assert(iov[n].iov_len == 1); assert(flags[n] & VRING_DESC_F_WRITE); /* * XXX * The guest should not be setting the BARRIER flag because * we don't advertise the capability. */ type = vbh->vbh_type & ~VBH_FLAG_BARRIER; writeop = (type == VBH_OP_WRITE); iolen = 0; for (i = 1; i < n; i++) { /* * - write op implies read-only descriptor, * - read/ident op implies write-only descriptor, * therefore test the inverse of the descriptor bit * to the op. */ assert(((flags[i] & VRING_DESC_F_WRITE) == 0) == writeop); iolen += iov[i].iov_len; } DPRINTF(("virtio-block: %s op, %d bytes, %d segs, offset %ld\n\r", writeop ? "write" : "read/ident", iolen, i - 1, offset)); switch (type) { case VBH_OP_READ: err = blockif_read(sc->bc, &io->io_req); break; case VBH_OP_WRITE: err = blockif_write(sc->bc, &io->io_req); break; case VBH_OP_FLUSH: case VBH_OP_FLUSH_OUT: err = blockif_flush(sc->bc, &io->io_req); break; case VBH_OP_IDENT: /* Assume a single buffer */ /* S/n equal to buffer is not zero-terminated. */ memset(iov[1].iov_base, 0, iov[1].iov_len); strncpy(iov[1].iov_base, sc->vbsc_ident, MIN(iov[1].iov_len, sizeof(sc->vbsc_ident))); pci_vtblk_done(&io->io_req, 0); return; default: pci_vtblk_done(&io->io_req, EOPNOTSUPP); return; } assert(err == 0); } static void pci_vtblk_notify(void *vsc, struct vqueue_info *vq) { struct pci_vtblk_softc *sc = vsc; while (vq_has_descs(vq)) pci_vtblk_proc(sc, vq); } static int pci_vtblk_init(struct vmctx *ctx, struct pci_devinst *pi, char *opts) { char bident[sizeof("XX:X:X")]; struct blockif_ctxt *bctxt; MD5_CTX mdctx; u_char digest[16]; struct pci_vtblk_softc *sc; off_t size; int i, sectsz, sts, sto; if (opts == NULL) { printf("virtio-block: backing device required\n"); return (1); } /* * The supplied backing file has to exist */ snprintf(bident, sizeof(bident), "%d:%d", pi->pi_slot, pi->pi_func); bctxt = blockif_open(opts, bident); if (bctxt == NULL) { perror("Could not open backing file"); return (1); } size = blockif_size(bctxt); sectsz = blockif_sectsz(bctxt); blockif_psectsz(bctxt, &sts, &sto); sc = calloc(1, sizeof(struct pci_vtblk_softc)); sc->bc = bctxt; for (i = 0; i < VTBLK_RINGSZ; i++) { struct pci_vtblk_ioreq *io = &sc->vbsc_ios[i]; io->io_req.br_callback = pci_vtblk_done; io->io_req.br_param = io; io->io_sc = sc; io->io_idx = i; } pthread_mutex_init(&sc->vsc_mtx, NULL); /* init virtio softc and virtqueues */ vi_softc_linkup(&sc->vbsc_vs, &vtblk_vi_consts, sc, pi, &sc->vbsc_vq); sc->vbsc_vs.vs_mtx = &sc->vsc_mtx; sc->vbsc_vq.vq_qsize = VTBLK_RINGSZ; /* sc->vbsc_vq.vq_notify = we have no per-queue notify */ /* * Create an identifier for the backing file. Use parts of the * md5 sum of the filename */ MD5Init(&mdctx); MD5Update(&mdctx, opts, strlen(opts)); MD5Final(digest, &mdctx); sprintf(sc->vbsc_ident, "BHYVE-%02X%02X-%02X%02X-%02X%02X", digest[0], digest[1], digest[2], digest[3], digest[4], digest[5]); /* setup virtio block config space */ sc->vbsc_cfg.vbc_capacity = size / DEV_BSIZE; /* 512-byte units */ sc->vbsc_cfg.vbc_size_max = 0; /* not negotiated */ sc->vbsc_cfg.vbc_seg_max = BLOCKIF_IOV_MAX; sc->vbsc_cfg.vbc_geometry.cylinders = 0; /* no geometry */ sc->vbsc_cfg.vbc_geometry.heads = 0; sc->vbsc_cfg.vbc_geometry.sectors = 0; sc->vbsc_cfg.vbc_blk_size = sectsz; sc->vbsc_cfg.vbc_topology.physical_block_exp = (sts > sectsz) ? (ffsll(sts / sectsz) - 1) : 0; sc->vbsc_cfg.vbc_topology.alignment_offset = (sto != 0) ? ((sts - sto) / sectsz) : 0; sc->vbsc_cfg.vbc_topology.min_io_size = 0; sc->vbsc_cfg.vbc_topology.opt_io_size = 0; sc->vbsc_cfg.vbc_writeback = 0; /* * Should we move some of this into virtio.c? Could * have the device, class, and subdev_0 as fields in * the virtio constants structure. */ pci_set_cfgdata16(pi, PCIR_DEVICE, VIRTIO_DEV_BLOCK); pci_set_cfgdata16(pi, PCIR_VENDOR, VIRTIO_VENDOR); pci_set_cfgdata8(pi, PCIR_CLASS, PCIC_STORAGE); pci_set_cfgdata16(pi, PCIR_SUBDEV_0, VIRTIO_TYPE_BLOCK); - pci_lintr_request(pi); - if (vi_intr_init(&sc->vbsc_vs, 1, fbsdrun_virtio_msix())) { blockif_close(sc->bc); free(sc); return (1); } vi_set_io_bar(&sc->vbsc_vs, 0); return (0); } static int pci_vtblk_cfgwrite(void *vsc, int offset, int size, uint32_t value) { DPRINTF(("vtblk: write to readonly reg %d\n\r", offset)); return (1); } static int pci_vtblk_cfgread(void *vsc, int offset, int size, uint32_t *retval) { struct pci_vtblk_softc *sc = vsc; void *ptr; /* our caller has already verified offset and size */ ptr = (uint8_t *)&sc->vbsc_cfg + offset; memcpy(retval, ptr, size); return (0); } struct pci_devemu pci_de_vblk = { .pe_emu = "virtio-blk", .pe_init = pci_vtblk_init, .pe_barwrite = vi_pci_write, .pe_barread = vi_pci_read }; PCI_EMUL_SET(pci_de_vblk); Index: head/usr.sbin/bhyve/pci_virtio_net.c =================================================================== --- head/usr.sbin/bhyve/pci_virtio_net.c (revision 280724) +++ head/usr.sbin/bhyve/pci_virtio_net.c (revision 280725) @@ -1,730 +1,728 @@ /*- * Copyright (c) 2011 NetApp, Inc. * All rights reserved. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY NETAPP, INC ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL NETAPP, INC OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * $FreeBSD$ */ #include __FBSDID("$FreeBSD$"); #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include "bhyverun.h" #include "pci_emul.h" #include "mevent.h" #include "virtio.h" #define VTNET_RINGSZ 1024 #define VTNET_MAXSEGS 32 /* * Host capabilities. Note that we only offer a few of these. */ #define VIRTIO_NET_F_CSUM (1 << 0) /* host handles partial cksum */ #define VIRTIO_NET_F_GUEST_CSUM (1 << 1) /* guest handles partial cksum */ #define VIRTIO_NET_F_MAC (1 << 5) /* host supplies MAC */ #define VIRTIO_NET_F_GSO_DEPREC (1 << 6) /* deprecated: host handles GSO */ #define VIRTIO_NET_F_GUEST_TSO4 (1 << 7) /* guest can rcv TSOv4 */ #define VIRTIO_NET_F_GUEST_TSO6 (1 << 8) /* guest can rcv TSOv6 */ #define VIRTIO_NET_F_GUEST_ECN (1 << 9) /* guest can rcv TSO with ECN */ #define VIRTIO_NET_F_GUEST_UFO (1 << 10) /* guest can rcv UFO */ #define VIRTIO_NET_F_HOST_TSO4 (1 << 11) /* host can rcv TSOv4 */ #define VIRTIO_NET_F_HOST_TSO6 (1 << 12) /* host can rcv TSOv6 */ #define VIRTIO_NET_F_HOST_ECN (1 << 13) /* host can rcv TSO with ECN */ #define VIRTIO_NET_F_HOST_UFO (1 << 14) /* host can rcv UFO */ #define VIRTIO_NET_F_MRG_RXBUF (1 << 15) /* host can merge RX buffers */ #define VIRTIO_NET_F_STATUS (1 << 16) /* config status field available */ #define VIRTIO_NET_F_CTRL_VQ (1 << 17) /* control channel available */ #define VIRTIO_NET_F_CTRL_RX (1 << 18) /* control channel RX mode support */ #define VIRTIO_NET_F_CTRL_VLAN (1 << 19) /* control channel VLAN filtering */ #define VIRTIO_NET_F_GUEST_ANNOUNCE \ (1 << 21) /* guest can send gratuitous pkts */ #define VTNET_S_HOSTCAPS \ ( VIRTIO_NET_F_MAC | VIRTIO_NET_F_MRG_RXBUF | VIRTIO_NET_F_STATUS | \ VIRTIO_F_NOTIFY_ON_EMPTY) /* * PCI config-space "registers" */ struct virtio_net_config { uint8_t mac[6]; uint16_t status; } __packed; /* * Queue definitions. */ #define VTNET_RXQ 0 #define VTNET_TXQ 1 #define VTNET_CTLQ 2 /* NB: not yet supported */ #define VTNET_MAXQ 3 /* * Fixed network header size */ struct virtio_net_rxhdr { uint8_t vrh_flags; uint8_t vrh_gso_type; uint16_t vrh_hdr_len; uint16_t vrh_gso_size; uint16_t vrh_csum_start; uint16_t vrh_csum_offset; uint16_t vrh_bufs; } __packed; /* * Debug printf */ static int pci_vtnet_debug; #define DPRINTF(params) if (pci_vtnet_debug) printf params #define WPRINTF(params) printf params /* * Per-device softc */ struct pci_vtnet_softc { struct virtio_softc vsc_vs; struct vqueue_info vsc_queues[VTNET_MAXQ - 1]; pthread_mutex_t vsc_mtx; struct mevent *vsc_mevp; int vsc_tapfd; int vsc_rx_ready; volatile int resetting; /* set and checked outside lock */ uint64_t vsc_features; /* negotiated features */ struct virtio_net_config vsc_config; pthread_mutex_t rx_mtx; int rx_in_progress; int rx_vhdrlen; int rx_merge; /* merged rx bufs in use */ pthread_t tx_tid; pthread_mutex_t tx_mtx; pthread_cond_t tx_cond; int tx_in_progress; }; static void pci_vtnet_reset(void *); /* static void pci_vtnet_notify(void *, struct vqueue_info *); */ static int pci_vtnet_cfgread(void *, int, int, uint32_t *); static int pci_vtnet_cfgwrite(void *, int, int, uint32_t); static void pci_vtnet_neg_features(void *, uint64_t); static struct virtio_consts vtnet_vi_consts = { "vtnet", /* our name */ VTNET_MAXQ - 1, /* we currently support 2 virtqueues */ sizeof(struct virtio_net_config), /* config reg size */ pci_vtnet_reset, /* reset */ NULL, /* device-wide qnotify -- not used */ pci_vtnet_cfgread, /* read PCI config */ pci_vtnet_cfgwrite, /* write PCI config */ pci_vtnet_neg_features, /* apply negotiated features */ VTNET_S_HOSTCAPS, /* our capabilities */ }; /* * If the transmit thread is active then stall until it is done. */ static void pci_vtnet_txwait(struct pci_vtnet_softc *sc) { pthread_mutex_lock(&sc->tx_mtx); while (sc->tx_in_progress) { pthread_mutex_unlock(&sc->tx_mtx); usleep(10000); pthread_mutex_lock(&sc->tx_mtx); } pthread_mutex_unlock(&sc->tx_mtx); } /* * If the receive thread is active then stall until it is done. */ static void pci_vtnet_rxwait(struct pci_vtnet_softc *sc) { pthread_mutex_lock(&sc->rx_mtx); while (sc->rx_in_progress) { pthread_mutex_unlock(&sc->rx_mtx); usleep(10000); pthread_mutex_lock(&sc->rx_mtx); } pthread_mutex_unlock(&sc->rx_mtx); } static void pci_vtnet_reset(void *vsc) { struct pci_vtnet_softc *sc = vsc; DPRINTF(("vtnet: device reset requested !\n")); sc->resetting = 1; /* * Wait for the transmit and receive threads to finish their * processing. */ pci_vtnet_txwait(sc); pci_vtnet_rxwait(sc); sc->vsc_rx_ready = 0; sc->rx_merge = 1; sc->rx_vhdrlen = sizeof(struct virtio_net_rxhdr); /* now reset rings, MSI-X vectors, and negotiated capabilities */ vi_reset_dev(&sc->vsc_vs); sc->resetting = 0; } /* * Called to send a buffer chain out to the tap device */ static void pci_vtnet_tap_tx(struct pci_vtnet_softc *sc, struct iovec *iov, int iovcnt, int len) { static char pad[60]; /* all zero bytes */ if (sc->vsc_tapfd == -1) return; /* * If the length is < 60, pad out to that and add the * extra zero'd segment to the iov. It is guaranteed that * there is always an extra iov available by the caller. */ if (len < 60) { iov[iovcnt].iov_base = pad; iov[iovcnt].iov_len = 60 - len; iovcnt++; } (void) writev(sc->vsc_tapfd, iov, iovcnt); } /* * Called when there is read activity on the tap file descriptor. * Each buffer posted by the guest is assumed to be able to contain * an entire ethernet frame + rx header. * MP note: the dummybuf is only used for discarding frames, so there * is no need for it to be per-vtnet or locked. */ static uint8_t dummybuf[2048]; static __inline struct iovec * rx_iov_trim(struct iovec *iov, int *niov, int tlen) { struct iovec *riov; /* XXX short-cut: assume first segment is >= tlen */ assert(iov[0].iov_len >= tlen); iov[0].iov_len -= tlen; if (iov[0].iov_len == 0) { assert(*niov > 1); *niov -= 1; riov = &iov[1]; } else { iov[0].iov_base = (void *)((uintptr_t)iov[0].iov_base + tlen); riov = &iov[0]; } return (riov); } static void pci_vtnet_tap_rx(struct pci_vtnet_softc *sc) { struct iovec iov[VTNET_MAXSEGS], *riov; struct vqueue_info *vq; void *vrx; int len, n; uint16_t idx; /* * Should never be called without a valid tap fd */ assert(sc->vsc_tapfd != -1); /* * But, will be called when the rx ring hasn't yet * been set up or the guest is resetting the device. */ if (!sc->vsc_rx_ready || sc->resetting) { /* * Drop the packet and try later. */ (void) read(sc->vsc_tapfd, dummybuf, sizeof(dummybuf)); return; } /* * Check for available rx buffers */ vq = &sc->vsc_queues[VTNET_RXQ]; if (!vq_has_descs(vq)) { /* * Drop the packet and try later. Interrupt on * empty, if that's negotiated. */ (void) read(sc->vsc_tapfd, dummybuf, sizeof(dummybuf)); vq_endchains(vq, 1); return; } do { /* * Get descriptor chain. */ n = vq_getchain(vq, &idx, iov, VTNET_MAXSEGS, NULL); assert(n >= 1 && n <= VTNET_MAXSEGS); /* * Get a pointer to the rx header, and use the * data immediately following it for the packet buffer. */ vrx = iov[0].iov_base; riov = rx_iov_trim(iov, &n, sc->rx_vhdrlen); len = readv(sc->vsc_tapfd, riov, n); if (len < 0 && errno == EWOULDBLOCK) { /* * No more packets, but still some avail ring * entries. Interrupt if needed/appropriate. */ vq_retchain(vq); vq_endchains(vq, 0); return; } /* * The only valid field in the rx packet header is the * number of buffers if merged rx bufs were negotiated. */ memset(vrx, 0, sc->rx_vhdrlen); if (sc->rx_merge) { struct virtio_net_rxhdr *vrxh; vrxh = vrx; vrxh->vrh_bufs = 1; } /* * Release this chain and handle more chains. */ vq_relchain(vq, idx, len + sc->rx_vhdrlen); } while (vq_has_descs(vq)); /* Interrupt if needed, including for NOTIFY_ON_EMPTY. */ vq_endchains(vq, 1); } static void pci_vtnet_tap_callback(int fd, enum ev_type type, void *param) { struct pci_vtnet_softc *sc = param; pthread_mutex_lock(&sc->rx_mtx); sc->rx_in_progress = 1; pci_vtnet_tap_rx(sc); sc->rx_in_progress = 0; pthread_mutex_unlock(&sc->rx_mtx); } static void pci_vtnet_ping_rxq(void *vsc, struct vqueue_info *vq) { struct pci_vtnet_softc *sc = vsc; /* * A qnotify means that the rx process can now begin */ if (sc->vsc_rx_ready == 0) { sc->vsc_rx_ready = 1; } } static void pci_vtnet_proctx(struct pci_vtnet_softc *sc, struct vqueue_info *vq) { struct iovec iov[VTNET_MAXSEGS + 1]; int i, n; int plen, tlen; uint16_t idx; /* * Obtain chain of descriptors. The first one is * really the header descriptor, so we need to sum * up two lengths: packet length and transfer length. */ n = vq_getchain(vq, &idx, iov, VTNET_MAXSEGS, NULL); assert(n >= 1 && n <= VTNET_MAXSEGS); plen = 0; tlen = iov[0].iov_len; for (i = 1; i < n; i++) { plen += iov[i].iov_len; tlen += iov[i].iov_len; } DPRINTF(("virtio: packet send, %d bytes, %d segs\n\r", plen, n)); pci_vtnet_tap_tx(sc, &iov[1], n - 1, plen); /* chain is processed, release it and set tlen */ vq_relchain(vq, idx, tlen); } static void pci_vtnet_ping_txq(void *vsc, struct vqueue_info *vq) { struct pci_vtnet_softc *sc = vsc; /* * Any ring entries to process? */ if (!vq_has_descs(vq)) return; /* Signal the tx thread for processing */ pthread_mutex_lock(&sc->tx_mtx); if (sc->tx_in_progress == 0) pthread_cond_signal(&sc->tx_cond); pthread_mutex_unlock(&sc->tx_mtx); } /* * Thread which will handle processing of TX desc */ static void * pci_vtnet_tx_thread(void *param) { struct pci_vtnet_softc *sc = param; struct vqueue_info *vq; int have_work, error; vq = &sc->vsc_queues[VTNET_TXQ]; /* * Let us wait till the tx queue pointers get initialised & * first tx signaled */ pthread_mutex_lock(&sc->tx_mtx); error = pthread_cond_wait(&sc->tx_cond, &sc->tx_mtx); assert(error == 0); for (;;) { /* note - tx mutex is locked here */ do { if (sc->resetting) have_work = 0; else have_work = vq_has_descs(vq); if (!have_work) { sc->tx_in_progress = 0; error = pthread_cond_wait(&sc->tx_cond, &sc->tx_mtx); assert(error == 0); } } while (!have_work); sc->tx_in_progress = 1; pthread_mutex_unlock(&sc->tx_mtx); do { /* * Run through entries, placing them into * iovecs and sending when an end-of-packet * is found */ pci_vtnet_proctx(sc, vq); } while (vq_has_descs(vq)); /* * Generate an interrupt if needed. */ vq_endchains(vq, 1); pthread_mutex_lock(&sc->tx_mtx); } } #ifdef notyet static void pci_vtnet_ping_ctlq(void *vsc, struct vqueue_info *vq) { DPRINTF(("vtnet: control qnotify!\n\r")); } #endif static int pci_vtnet_parsemac(char *mac_str, uint8_t *mac_addr) { struct ether_addr *ea; char *tmpstr; char zero_addr[ETHER_ADDR_LEN] = { 0, 0, 0, 0, 0, 0 }; tmpstr = strsep(&mac_str,"="); if ((mac_str != NULL) && (!strcmp(tmpstr,"mac"))) { ea = ether_aton(mac_str); if (ea == NULL || ETHER_IS_MULTICAST(ea->octet) || memcmp(ea->octet, zero_addr, ETHER_ADDR_LEN) == 0) { fprintf(stderr, "Invalid MAC %s\n", mac_str); return (EINVAL); } else memcpy(mac_addr, ea->octet, ETHER_ADDR_LEN); } return (0); } static int pci_vtnet_init(struct vmctx *ctx, struct pci_devinst *pi, char *opts) { MD5_CTX mdctx; unsigned char digest[16]; char nstr[80]; char tname[MAXCOMLEN + 1]; struct pci_vtnet_softc *sc; char *devname; char *vtopts; int mac_provided; sc = calloc(1, sizeof(struct pci_vtnet_softc)); pthread_mutex_init(&sc->vsc_mtx, NULL); vi_softc_linkup(&sc->vsc_vs, &vtnet_vi_consts, sc, pi, sc->vsc_queues); sc->vsc_vs.vs_mtx = &sc->vsc_mtx; sc->vsc_queues[VTNET_RXQ].vq_qsize = VTNET_RINGSZ; sc->vsc_queues[VTNET_RXQ].vq_notify = pci_vtnet_ping_rxq; sc->vsc_queues[VTNET_TXQ].vq_qsize = VTNET_RINGSZ; sc->vsc_queues[VTNET_TXQ].vq_notify = pci_vtnet_ping_txq; #ifdef notyet sc->vsc_queues[VTNET_CTLQ].vq_qsize = VTNET_RINGSZ; sc->vsc_queues[VTNET_CTLQ].vq_notify = pci_vtnet_ping_ctlq; #endif /* * Attempt to open the tap device and read the MAC address * if specified */ mac_provided = 0; sc->vsc_tapfd = -1; if (opts != NULL) { char tbuf[80]; int err; devname = vtopts = strdup(opts); (void) strsep(&vtopts, ","); if (vtopts != NULL) { err = pci_vtnet_parsemac(vtopts, sc->vsc_config.mac); if (err != 0) { free(devname); return (err); } mac_provided = 1; } strcpy(tbuf, "/dev/"); strlcat(tbuf, devname, sizeof(tbuf)); free(devname); sc->vsc_tapfd = open(tbuf, O_RDWR); if (sc->vsc_tapfd == -1) { WPRINTF(("open of tap device %s failed\n", tbuf)); } else { /* * Set non-blocking and register for read * notifications with the event loop */ int opt = 1; if (ioctl(sc->vsc_tapfd, FIONBIO, &opt) < 0) { WPRINTF(("tap device O_NONBLOCK failed\n")); close(sc->vsc_tapfd); sc->vsc_tapfd = -1; } sc->vsc_mevp = mevent_add(sc->vsc_tapfd, EVF_READ, pci_vtnet_tap_callback, sc); if (sc->vsc_mevp == NULL) { WPRINTF(("Could not register event\n")); close(sc->vsc_tapfd); sc->vsc_tapfd = -1; } } } /* * The default MAC address is the standard NetApp OUI of 00-a0-98, * followed by an MD5 of the PCI slot/func number and dev name */ if (!mac_provided) { snprintf(nstr, sizeof(nstr), "%d-%d-%s", pi->pi_slot, pi->pi_func, vmname); MD5Init(&mdctx); MD5Update(&mdctx, nstr, strlen(nstr)); MD5Final(digest, &mdctx); sc->vsc_config.mac[0] = 0x00; sc->vsc_config.mac[1] = 0xa0; sc->vsc_config.mac[2] = 0x98; sc->vsc_config.mac[3] = digest[0]; sc->vsc_config.mac[4] = digest[1]; sc->vsc_config.mac[5] = digest[2]; } /* initialize config space */ pci_set_cfgdata16(pi, PCIR_DEVICE, VIRTIO_DEV_NET); pci_set_cfgdata16(pi, PCIR_VENDOR, VIRTIO_VENDOR); pci_set_cfgdata8(pi, PCIR_CLASS, PCIC_NETWORK); pci_set_cfgdata16(pi, PCIR_SUBDEV_0, VIRTIO_TYPE_NET); - pci_lintr_request(pi); - /* link always up */ sc->vsc_config.status = 1; /* use BAR 1 to map MSI-X table and PBA, if we're using MSI-X */ if (vi_intr_init(&sc->vsc_vs, 1, fbsdrun_virtio_msix())) return (1); /* use BAR 0 to map config regs in IO space */ vi_set_io_bar(&sc->vsc_vs, 0); sc->resetting = 0; sc->rx_merge = 1; sc->rx_vhdrlen = sizeof(struct virtio_net_rxhdr); sc->rx_in_progress = 0; pthread_mutex_init(&sc->rx_mtx, NULL); /* * Initialize tx semaphore & spawn TX processing thread. * As of now, only one thread for TX desc processing is * spawned. */ sc->tx_in_progress = 0; pthread_mutex_init(&sc->tx_mtx, NULL); pthread_cond_init(&sc->tx_cond, NULL); pthread_create(&sc->tx_tid, NULL, pci_vtnet_tx_thread, (void *)sc); snprintf(tname, sizeof(tname), "vtnet-%d:%d tx", pi->pi_slot, pi->pi_func); pthread_set_name_np(sc->tx_tid, tname); return (0); } static int pci_vtnet_cfgwrite(void *vsc, int offset, int size, uint32_t value) { struct pci_vtnet_softc *sc = vsc; void *ptr; if (offset < 6) { assert(offset + size <= 6); /* * The driver is allowed to change the MAC address */ ptr = &sc->vsc_config.mac[offset]; memcpy(ptr, &value, size); } else { /* silently ignore other writes */ DPRINTF(("vtnet: write to readonly reg %d\n\r", offset)); } return (0); } static int pci_vtnet_cfgread(void *vsc, int offset, int size, uint32_t *retval) { struct pci_vtnet_softc *sc = vsc; void *ptr; ptr = (uint8_t *)&sc->vsc_config + offset; memcpy(retval, ptr, size); return (0); } static void pci_vtnet_neg_features(void *vsc, uint64_t negotiated_features) { struct pci_vtnet_softc *sc = vsc; sc->vsc_features = negotiated_features; if (!(sc->vsc_features & VIRTIO_NET_F_MRG_RXBUF)) { sc->rx_merge = 0; /* non-merge rx header is 2 bytes shorter */ sc->rx_vhdrlen -= 2; } } struct pci_devemu pci_de_vnet = { .pe_emu = "virtio-net", .pe_init = pci_vtnet_init, .pe_barwrite = vi_pci_write, .pe_barread = vi_pci_read }; PCI_EMUL_SET(pci_de_vnet); Index: head/usr.sbin/bhyve/virtio.c =================================================================== --- head/usr.sbin/bhyve/virtio.c (revision 280724) +++ head/usr.sbin/bhyve/virtio.c (revision 280725) @@ -1,772 +1,777 @@ /*- * Copyright (c) 2013 Chris Torek * All rights reserved. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. */ #include __FBSDID("$FreeBSD$"); #include #include #include #include #include #include #include "bhyverun.h" #include "pci_emul.h" #include "virtio.h" /* * Functions for dealing with generalized "virtual devices" as * defined by */ /* * In case we decide to relax the "virtio softc comes at the * front of virtio-based device softc" constraint, let's use * this to convert. */ #define DEV_SOFTC(vs) ((void *)(vs)) /* * Link a virtio_softc to its constants, the device softc, and * the PCI emulation. */ void vi_softc_linkup(struct virtio_softc *vs, struct virtio_consts *vc, void *dev_softc, struct pci_devinst *pi, struct vqueue_info *queues) { int i; /* vs and dev_softc addresses must match */ assert((void *)vs == dev_softc); vs->vs_vc = vc; vs->vs_pi = pi; pi->pi_arg = vs; vs->vs_queues = queues; for (i = 0; i < vc->vc_nvq; i++) { queues[i].vq_vs = vs; queues[i].vq_num = i; } } /* * Reset device (device-wide). This erases all queues, i.e., * all the queues become invalid (though we don't wipe out the * internal pointers, we just clear the VQ_ALLOC flag). * * It resets negotiated features to "none". * * If MSI-X is enabled, this also resets all the vectors to NO_VECTOR. */ void vi_reset_dev(struct virtio_softc *vs) { struct vqueue_info *vq; int i, nvq; if (vs->vs_mtx) assert(pthread_mutex_isowned_np(vs->vs_mtx)); nvq = vs->vs_vc->vc_nvq; for (vq = vs->vs_queues, i = 0; i < nvq; vq++, i++) { vq->vq_flags = 0; vq->vq_last_avail = 0; vq->vq_save_used = 0; vq->vq_pfn = 0; vq->vq_msix_idx = VIRTIO_MSI_NO_VECTOR; } vs->vs_negotiated_caps = 0; vs->vs_curq = 0; /* vs->vs_status = 0; -- redundant */ if (vs->vs_isr) pci_lintr_deassert(vs->vs_pi); vs->vs_isr = 0; vs->vs_msix_cfg_idx = VIRTIO_MSI_NO_VECTOR; } /* * Set I/O BAR (usually 0) to map PCI config registers. */ void vi_set_io_bar(struct virtio_softc *vs, int barnum) { size_t size; /* * ??? should we use CFG0 if MSI-X is disabled? * Existing code did not... */ size = VTCFG_R_CFG1 + vs->vs_vc->vc_cfgsize; pci_emul_alloc_bar(vs->vs_pi, barnum, PCIBAR_IO, size); } /* * Initialize MSI-X vector capabilities if we're to use MSI-X, * or MSI capabilities if not. * * We assume we want one MSI-X vector per queue, here, plus one * for the config vec. */ int vi_intr_init(struct virtio_softc *vs, int barnum, int use_msix) { int nvec; if (use_msix) { vs->vs_flags |= VIRTIO_USE_MSIX; VS_LOCK(vs); vi_reset_dev(vs); /* set all vectors to NO_VECTOR */ VS_UNLOCK(vs); nvec = vs->vs_vc->vc_nvq + 1; if (pci_emul_add_msixcap(vs->vs_pi, nvec, barnum)) return (1); } else vs->vs_flags &= ~VIRTIO_USE_MSIX; + /* Only 1 MSI vector for bhyve */ pci_emul_add_msicap(vs->vs_pi, 1); + + /* Legacy interrupts are mandatory for virtio devices */ + pci_lintr_request(vs->vs_pi); + return (0); } /* * Initialize the currently-selected virtio queue (vs->vs_curq). * The guest just gave us a page frame number, from which we can * calculate the addresses of the queue. */ void vi_vq_init(struct virtio_softc *vs, uint32_t pfn) { struct vqueue_info *vq; uint64_t phys; size_t size; char *base; vq = &vs->vs_queues[vs->vs_curq]; vq->vq_pfn = pfn; phys = (uint64_t)pfn << VRING_PFN; size = vring_size(vq->vq_qsize); base = paddr_guest2host(vs->vs_pi->pi_vmctx, phys, size); /* First page(s) are descriptors... */ vq->vq_desc = (struct virtio_desc *)base; base += vq->vq_qsize * sizeof(struct virtio_desc); /* ... immediately followed by "avail" ring (entirely uint16_t's) */ vq->vq_avail = (struct vring_avail *)base; base += (2 + vq->vq_qsize + 1) * sizeof(uint16_t); /* Then it's rounded up to the next page... */ base = (char *)roundup2((uintptr_t)base, VRING_ALIGN); /* ... and the last page(s) are the used ring. */ vq->vq_used = (struct vring_used *)base; /* Mark queue as allocated, and start at 0 when we use it. */ vq->vq_flags = VQ_ALLOC; vq->vq_last_avail = 0; vq->vq_save_used = 0; } /* * Helper inline for vq_getchain(): record the i'th "real" * descriptor. */ static inline void _vq_record(int i, volatile struct virtio_desc *vd, struct vmctx *ctx, struct iovec *iov, int n_iov, uint16_t *flags) { if (i >= n_iov) return; iov[i].iov_base = paddr_guest2host(ctx, vd->vd_addr, vd->vd_len); iov[i].iov_len = vd->vd_len; if (flags != NULL) flags[i] = vd->vd_flags; } #define VQ_MAX_DESCRIPTORS 512 /* see below */ /* * Examine the chain of descriptors starting at the "next one" to * make sure that they describe a sensible request. If so, return * the number of "real" descriptors that would be needed/used in * acting on this request. This may be smaller than the number of * available descriptors, e.g., if there are two available but * they are two separate requests, this just returns 1. Or, it * may be larger: if there are indirect descriptors involved, * there may only be one descriptor available but it may be an * indirect pointing to eight more. We return 8 in this case, * i.e., we do not count the indirect descriptors, only the "real" * ones. * * Basically, this vets the vd_flags and vd_next field of each * descriptor and tells you how many are involved. Since some may * be indirect, this also needs the vmctx (in the pci_devinst * at vs->vs_pi) so that it can find indirect descriptors. * * As we process each descriptor, we copy and adjust it (guest to * host address wise, also using the vmtctx) into the given iov[] * array (of the given size). If the array overflows, we stop * placing values into the array but keep processing descriptors, * up to VQ_MAX_DESCRIPTORS, before giving up and returning -1. * So you, the caller, must not assume that iov[] is as big as the * return value (you can process the same thing twice to allocate * a larger iov array if needed, or supply a zero length to find * out how much space is needed). * * If you want to verify the WRITE flag on each descriptor, pass a * non-NULL "flags" pointer to an array of "uint16_t" of the same size * as n_iov and we'll copy each vd_flags field after unwinding any * indirects. * * If some descriptor(s) are invalid, this prints a diagnostic message * and returns -1. If no descriptors are ready now it simply returns 0. * * You are assumed to have done a vq_ring_ready() if needed (note * that vq_has_descs() does one). */ int vq_getchain(struct vqueue_info *vq, uint16_t *pidx, struct iovec *iov, int n_iov, uint16_t *flags) { int i; u_int ndesc, n_indir; u_int idx, next; volatile struct virtio_desc *vdir, *vindir, *vp; struct vmctx *ctx; struct virtio_softc *vs; const char *name; vs = vq->vq_vs; name = vs->vs_vc->vc_name; /* * Note: it's the responsibility of the guest not to * update vq->vq_avail->va_idx until all of the descriptors * the guest has written are valid (including all their * vd_next fields and vd_flags). * * Compute (last_avail - va_idx) in integers mod 2**16. This is * the number of descriptors the device has made available * since the last time we updated vq->vq_last_avail. * * We just need to do the subtraction as an unsigned int, * then trim off excess bits. */ idx = vq->vq_last_avail; ndesc = (uint16_t)((u_int)vq->vq_avail->va_idx - idx); if (ndesc == 0) return (0); if (ndesc > vq->vq_qsize) { /* XXX need better way to diagnose issues */ fprintf(stderr, "%s: ndesc (%u) out of range, driver confused?\r\n", name, (u_int)ndesc); return (-1); } /* * Now count/parse "involved" descriptors starting from * the head of the chain. * * To prevent loops, we could be more complicated and * check whether we're re-visiting a previously visited * index, but we just abort if the count gets excessive. */ ctx = vs->vs_pi->pi_vmctx; *pidx = next = vq->vq_avail->va_ring[idx & (vq->vq_qsize - 1)]; vq->vq_last_avail++; for (i = 0; i < VQ_MAX_DESCRIPTORS; next = vdir->vd_next) { if (next >= vq->vq_qsize) { fprintf(stderr, "%s: descriptor index %u out of range, " "driver confused?\r\n", name, next); return (-1); } vdir = &vq->vq_desc[next]; if ((vdir->vd_flags & VRING_DESC_F_INDIRECT) == 0) { _vq_record(i, vdir, ctx, iov, n_iov, flags); i++; } else if ((vs->vs_negotiated_caps & VIRTIO_RING_F_INDIRECT_DESC) == 0) { fprintf(stderr, "%s: descriptor has forbidden INDIRECT flag, " "driver confused?\r\n", name); return (-1); } else { n_indir = vdir->vd_len / 16; if ((vdir->vd_len & 0xf) || n_indir == 0) { fprintf(stderr, "%s: invalid indir len 0x%x, " "driver confused?\r\n", name, (u_int)vdir->vd_len); return (-1); } vindir = paddr_guest2host(ctx, vdir->vd_addr, vdir->vd_len); /* * Indirects start at the 0th, then follow * their own embedded "next"s until those run * out. Each one's indirect flag must be off * (we don't really have to check, could just * ignore errors...). */ next = 0; for (;;) { vp = &vindir[next]; if (vp->vd_flags & VRING_DESC_F_INDIRECT) { fprintf(stderr, "%s: indirect desc has INDIR flag," " driver confused?\r\n", name); return (-1); } _vq_record(i, vp, ctx, iov, n_iov, flags); if (++i > VQ_MAX_DESCRIPTORS) goto loopy; if ((vp->vd_flags & VRING_DESC_F_NEXT) == 0) break; next = vp->vd_next; if (next >= n_indir) { fprintf(stderr, "%s: invalid next %u > %u, " "driver confused?\r\n", name, (u_int)next, n_indir); return (-1); } } } if ((vdir->vd_flags & VRING_DESC_F_NEXT) == 0) return (i); } loopy: fprintf(stderr, "%s: descriptor loop? count > %d - driver confused?\r\n", name, i); return (-1); } /* * Return the currently-first request chain back to the available queue. * * (This chain is the one you handled when you called vq_getchain() * and used its positive return value.) */ void vq_retchain(struct vqueue_info *vq) { vq->vq_last_avail--; } /* * Return specified request chain to the guest, setting its I/O length * to the provided value. * * (This chain is the one you handled when you called vq_getchain() * and used its positive return value.) */ void vq_relchain(struct vqueue_info *vq, uint16_t idx, uint32_t iolen) { uint16_t uidx, mask; volatile struct vring_used *vuh; volatile struct virtio_used *vue; /* * Notes: * - mask is N-1 where N is a power of 2 so computes x % N * - vuh points to the "used" data shared with guest * - vue points to the "used" ring entry we want to update * - head is the same value we compute in vq_iovecs(). * * (I apologize for the two fields named vu_idx; the * virtio spec calls the one that vue points to, "id"...) */ mask = vq->vq_qsize - 1; vuh = vq->vq_used; uidx = vuh->vu_idx; vue = &vuh->vu_ring[uidx++ & mask]; vue->vu_idx = idx; vue->vu_tlen = iolen; vuh->vu_idx = uidx; } /* * Driver has finished processing "available" chains and calling * vq_relchain on each one. If driver used all the available * chains, used_all should be set. * * If the "used" index moved we may need to inform the guest, i.e., * deliver an interrupt. Even if the used index did NOT move we * may need to deliver an interrupt, if the avail ring is empty and * we are supposed to interrupt on empty. * * Note that used_all_avail is provided by the caller because it's * a snapshot of the ring state when he decided to finish interrupt * processing -- it's possible that descriptors became available after * that point. (It's also typically a constant 1/True as well.) */ void vq_endchains(struct vqueue_info *vq, int used_all_avail) { struct virtio_softc *vs; uint16_t event_idx, new_idx, old_idx; int intr; /* * Interrupt generation: if we're using EVENT_IDX, * interrupt if we've crossed the event threshold. * Otherwise interrupt is generated if we added "used" entries, * but suppressed by VRING_AVAIL_F_NO_INTERRUPT. * * In any case, though, if NOTIFY_ON_EMPTY is set and the * entire avail was processed, we need to interrupt always. */ vs = vq->vq_vs; old_idx = vq->vq_save_used; vq->vq_save_used = new_idx = vq->vq_used->vu_idx; if (used_all_avail && (vs->vs_negotiated_caps & VIRTIO_F_NOTIFY_ON_EMPTY)) intr = 1; else if (vs->vs_negotiated_caps & VIRTIO_RING_F_EVENT_IDX) { event_idx = VQ_USED_EVENT_IDX(vq); /* * This calculation is per docs and the kernel * (see src/sys/dev/virtio/virtio_ring.h). */ intr = (uint16_t)(new_idx - event_idx - 1) < (uint16_t)(new_idx - old_idx); } else { intr = new_idx != old_idx && !(vq->vq_avail->va_flags & VRING_AVAIL_F_NO_INTERRUPT); } if (intr) vq_interrupt(vs, vq); } /* Note: these are in sorted order to make for a fast search */ static struct config_reg { uint16_t cr_offset; /* register offset */ uint8_t cr_size; /* size (bytes) */ uint8_t cr_ro; /* true => reg is read only */ const char *cr_name; /* name of reg */ } config_regs[] = { { VTCFG_R_HOSTCAP, 4, 1, "HOSTCAP" }, { VTCFG_R_GUESTCAP, 4, 0, "GUESTCAP" }, { VTCFG_R_PFN, 4, 0, "PFN" }, { VTCFG_R_QNUM, 2, 1, "QNUM" }, { VTCFG_R_QSEL, 2, 0, "QSEL" }, { VTCFG_R_QNOTIFY, 2, 0, "QNOTIFY" }, { VTCFG_R_STATUS, 1, 0, "STATUS" }, { VTCFG_R_ISR, 1, 0, "ISR" }, { VTCFG_R_CFGVEC, 2, 0, "CFGVEC" }, { VTCFG_R_QVEC, 2, 0, "QVEC" }, }; static inline struct config_reg * vi_find_cr(int offset) { u_int hi, lo, mid; struct config_reg *cr; lo = 0; hi = sizeof(config_regs) / sizeof(*config_regs) - 1; while (hi >= lo) { mid = (hi + lo) >> 1; cr = &config_regs[mid]; if (cr->cr_offset == offset) return (cr); if (cr->cr_offset < offset) lo = mid + 1; else hi = mid - 1; } return (NULL); } /* * Handle pci config space reads. * If it's to the MSI-X info, do that. * If it's part of the virtio standard stuff, do that. * Otherwise dispatch to the actual driver. */ uint64_t vi_pci_read(struct vmctx *ctx, int vcpu, struct pci_devinst *pi, int baridx, uint64_t offset, int size) { struct virtio_softc *vs = pi->pi_arg; struct virtio_consts *vc; struct config_reg *cr; uint64_t virtio_config_size, max; const char *name; uint32_t newoff; uint32_t value; int error; if (vs->vs_flags & VIRTIO_USE_MSIX) { if (baridx == pci_msix_table_bar(pi) || baridx == pci_msix_pba_bar(pi)) { return (pci_emul_msix_tread(pi, offset, size)); } } /* XXX probably should do something better than just assert() */ assert(baridx == 0); if (vs->vs_mtx) pthread_mutex_lock(vs->vs_mtx); vc = vs->vs_vc; name = vc->vc_name; value = size == 1 ? 0xff : size == 2 ? 0xffff : 0xffffffff; if (size != 1 && size != 2 && size != 4) goto bad; if (pci_msix_enabled(pi)) virtio_config_size = VTCFG_R_CFG1; else virtio_config_size = VTCFG_R_CFG0; if (offset >= virtio_config_size) { /* * Subtract off the standard size (including MSI-X * registers if enabled) and dispatch to underlying driver. * If that fails, fall into general code. */ newoff = offset - virtio_config_size; max = vc->vc_cfgsize ? vc->vc_cfgsize : 0x100000000; if (newoff + size > max) goto bad; error = (*vc->vc_cfgread)(DEV_SOFTC(vs), newoff, size, &value); if (!error) goto done; } bad: cr = vi_find_cr(offset); if (cr == NULL || cr->cr_size != size) { if (cr != NULL) { /* offset must be OK, so size must be bad */ fprintf(stderr, "%s: read from %s: bad size %d\r\n", name, cr->cr_name, size); } else { fprintf(stderr, "%s: read from bad offset/size %jd/%d\r\n", name, (uintmax_t)offset, size); } goto done; } switch (offset) { case VTCFG_R_HOSTCAP: value = vc->vc_hv_caps; break; case VTCFG_R_GUESTCAP: value = vs->vs_negotiated_caps; break; case VTCFG_R_PFN: if (vs->vs_curq < vc->vc_nvq) value = vs->vs_queues[vs->vs_curq].vq_pfn; break; case VTCFG_R_QNUM: value = vs->vs_curq < vc->vc_nvq ? vs->vs_queues[vs->vs_curq].vq_qsize : 0; break; case VTCFG_R_QSEL: value = vs->vs_curq; break; case VTCFG_R_QNOTIFY: value = 0; /* XXX */ break; case VTCFG_R_STATUS: value = vs->vs_status; break; case VTCFG_R_ISR: value = vs->vs_isr; vs->vs_isr = 0; /* a read clears this flag */ if (value) pci_lintr_deassert(pi); break; case VTCFG_R_CFGVEC: value = vs->vs_msix_cfg_idx; break; case VTCFG_R_QVEC: value = vs->vs_curq < vc->vc_nvq ? vs->vs_queues[vs->vs_curq].vq_msix_idx : VIRTIO_MSI_NO_VECTOR; break; } done: if (vs->vs_mtx) pthread_mutex_unlock(vs->vs_mtx); return (value); } /* * Handle pci config space writes. * If it's to the MSI-X info, do that. * If it's part of the virtio standard stuff, do that. * Otherwise dispatch to the actual driver. */ void vi_pci_write(struct vmctx *ctx, int vcpu, struct pci_devinst *pi, int baridx, uint64_t offset, int size, uint64_t value) { struct virtio_softc *vs = pi->pi_arg; struct vqueue_info *vq; struct virtio_consts *vc; struct config_reg *cr; uint64_t virtio_config_size, max; const char *name; uint32_t newoff; int error; if (vs->vs_flags & VIRTIO_USE_MSIX) { if (baridx == pci_msix_table_bar(pi) || baridx == pci_msix_pba_bar(pi)) { pci_emul_msix_twrite(pi, offset, size, value); return; } } /* XXX probably should do something better than just assert() */ assert(baridx == 0); if (vs->vs_mtx) pthread_mutex_lock(vs->vs_mtx); vc = vs->vs_vc; name = vc->vc_name; if (size != 1 && size != 2 && size != 4) goto bad; if (pci_msix_enabled(pi)) virtio_config_size = VTCFG_R_CFG1; else virtio_config_size = VTCFG_R_CFG0; if (offset >= virtio_config_size) { /* * Subtract off the standard size (including MSI-X * registers if enabled) and dispatch to underlying driver. */ newoff = offset - virtio_config_size; max = vc->vc_cfgsize ? vc->vc_cfgsize : 0x100000000; if (newoff + size > max) goto bad; error = (*vc->vc_cfgwrite)(DEV_SOFTC(vs), newoff, size, value); if (!error) goto done; } bad: cr = vi_find_cr(offset); if (cr == NULL || cr->cr_size != size || cr->cr_ro) { if (cr != NULL) { /* offset must be OK, wrong size and/or reg is R/O */ if (cr->cr_size != size) fprintf(stderr, "%s: write to %s: bad size %d\r\n", name, cr->cr_name, size); if (cr->cr_ro) fprintf(stderr, "%s: write to read-only reg %s\r\n", name, cr->cr_name); } else { fprintf(stderr, "%s: write to bad offset/size %jd/%d\r\n", name, (uintmax_t)offset, size); } goto done; } switch (offset) { case VTCFG_R_GUESTCAP: vs->vs_negotiated_caps = value & vc->vc_hv_caps; if (vc->vc_apply_features) (*vc->vc_apply_features)(DEV_SOFTC(vs), vs->vs_negotiated_caps); break; case VTCFG_R_PFN: if (vs->vs_curq >= vc->vc_nvq) goto bad_qindex; vi_vq_init(vs, value); break; case VTCFG_R_QSEL: /* * Note that the guest is allowed to select an * invalid queue; we just need to return a QNUM * of 0 while the bad queue is selected. */ vs->vs_curq = value; break; case VTCFG_R_QNOTIFY: if (value >= vc->vc_nvq) { fprintf(stderr, "%s: queue %d notify out of range\r\n", name, (int)value); goto done; } vq = &vs->vs_queues[value]; if (vq->vq_notify) (*vq->vq_notify)(DEV_SOFTC(vs), vq); else if (vc->vc_qnotify) (*vc->vc_qnotify)(DEV_SOFTC(vs), vq); else fprintf(stderr, "%s: qnotify queue %d: missing vq/vc notify\r\n", name, (int)value); break; case VTCFG_R_STATUS: vs->vs_status = value; if (value == 0) (*vc->vc_reset)(DEV_SOFTC(vs)); break; case VTCFG_R_CFGVEC: vs->vs_msix_cfg_idx = value; break; case VTCFG_R_QVEC: if (vs->vs_curq >= vc->vc_nvq) goto bad_qindex; vq = &vs->vs_queues[vs->vs_curq]; vq->vq_msix_idx = value; break; } goto done; bad_qindex: fprintf(stderr, "%s: write config reg %s: curq %d >= max %d\r\n", name, cr->cr_name, vs->vs_curq, vc->vc_nvq); done: if (vs->vs_mtx) pthread_mutex_unlock(vs->vs_mtx); }